A WGAN-GP Approach for Data Imputation in Photovoltaic Power Prediction

The increasing adoption of photovoltaic (PV) systems has introduced challenges for grid stability due to the intermittent nature of PV power generation. Accurate forecasting and data quality are critical for effective integration into power grids. However, PV power records often contain missing data due to system downtime, posing difficulties for pattern recognition and model accuracy. To address this, we propose a GAN-based data imputation method tailored for PV power generation. Unlike traditional GANs used in image generation, our method ensures smooth transitions with existing data by utilizing a data-guided GAN framework with quasi-convex properties. To stabilize training, we introduce a gradient penalty mechanism and a single-batch multi-iteration strategy. Our contributions include analyzing the necessity of data imputation, designing a novel conditional GAN-based network for PV data generation, and validating the generated data using frequency domain analysis, t-NSE, and prediction performance. This approach significantly enhances data continuity and reliability in PV forecasting tasks.